Re: [PATCH v7 12/13] ext4: switch to multigrain timestamps

[Jeff Layton]

On Tue, 2023-09-19 at 13:04 +0200, Jan Kara wrote:
> On Tue 19-09-23 15:05:24, Xi Ruoyao wrote:
> > On Mon, 2023-08-07 at 15:38 -0400, Jeff Layton wrote:
> > > Enable multigrain timestamps, which should ensure that there is an
> > > apparent change to the timestamp whenever it has been written after
> > > being actively observed via getattr.
> > > 
> > > For ext4, we only need to enable the FS_MGTIME flag.
> > 
> > Hi Jeff,
> > 
> > This patch causes a gnulib test failure:
> > 
> > $ ~/sources/lfs/grep-3.11/gnulib-tests/test-stat-time
> > test-stat-time.c:141: assertion 'statinfo[0].st_mtime < statinfo[2].st_mtime || (statinfo[0].st_mtime == statinfo[2].st_mtime && (get_stat_mtime_ns (&statinfo[0]) < get_stat_mtime_ns (&statinfo[2])))' failed
> > Aborted (core dumped)
> > 
> > The source code of the test:
> > https://git.savannah.gnu.org/cgit/gnulib.git/tree/tests/test-stat-time.c
> > 
> > Is this an expected change?
> 
> Kind of yes. The test first tries to estimate filesystem timestamp
> granularity in nap() function - due to this patch, the detected granularity
> will likely be 1 ns so effectively all the test calls will happen
> immediately one after another. But we don't bother setting the timestamps
> with more than 1 jiffy (usually 4 ms) precision unless we think someone is
> watching. So as a result timestamps of all stamp1 and stamp2 files are
> going to be equal which makes the test fail.
> 

That was my take too. The multigrain ctime changes are probably causing
nap() to settle on too small a time delta.

> The ultimate problem is that a sequence like:
> 
> write(f1)
> stat(f2)
> write(f2)
> stat(f2)
> write(f1)
> stat(f1)
>
> can result in f1 timestamp to be (slightly) lower than the final f2
> timestamp because the second write to f1 didn't bother updating the
> timestamp. That can indeed be a bit confusing to programs if they compare
> timestamps between two files. Jeff?
> 

Basically yes. When there is no stat() call issued on the file in
between writes, the kernel will use coarse-grained timestamps when
updating it (since no one is watching).


I'm not sure what we can do for this test. The nap() function is making
an assumption that the timestamp granularity will be constant, and that
isn't necessarily the case now.

-- 
Jeff Layton <jlayton@xxxxxxxxxx>